METHOD FOR SEGMENTING UTTERANCES BY USING PARTNER'S RESPONSE

• US 20080154594A1
• Filed: 12/26/2007
• Published: 06/26/2008
• Est. Priority Date: 12/26/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Kurata, Gakuto, Itoh, Nobuyasu
• Granted Patent: US 8,793,132 B2
• Time in Patent Office: Days
• US Class Current: 704/235
• CPC Class Codes: G10L 15/04 Segmentation; Word boundary detection
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules

Abstract

An apparatus, method and program for dividing a conversational dialog into utterance. The apparatus includes: a computer processor; a word database for storing spellings and pronunciations of words; a grammar database for storing syntactic rules on words; a pause detecting section which detects a pause location in a channel making a main speech among conversational dialogs inputted in at least two channels; an acknowledgement detecting section which detects an acknowledgement location in a channel not making the main speech; a boundary-candidate extracting section which extracts boundary candidates in the main speech, by extracting pauses existing within a predetermined range before and after a base point that is the acknowledgement location; and a recognizing unit which outputs a word string of the main speech segmented by one of the extracted boundary candidates after dividing the segmented speech into optimal utterance in reference to the word database and grammar database.

Claims

1. An apparatus for dividing a conversational dialog into utterance, comprising:

a computer processor;

a word database for storing spellings and pronunciations of words;

a grammar database for storing syntactic rules on words;

a pause detecting section which detects a pause location in a channel making a main speech among conversational dialogs inputted in at least two channels;

an acknowledgement detecting section which detects an acknowledgement location in a channel not making the main speech;

a boundary-candidate extracting section which extracts boundary candidates in the main speech, by extracting pauses existing within a predetermined range before and after a base point that is the acknowledgement location; and

a recognizing unit which outputs a word string of the main speech segmented by one of the extracted boundary candidates after dividing the segmented speech into optimal utterance in reference to the word database and grammar database.

2. The apparatus according to claim 1, wherein the grammar database includes fixed-phrase grammar, acknowledgement grammar and recognition grammar.

3. The apparatus according to claim 2, comprising a recognition-target segment determination unit, wherein:

the fixed-phrase grammar includes fixed phrases for starting and ending a confirmation,

the word database includes spellings and pronunciations of the fixed phrases for starting and ending a confirmation,

the recognition-target segment determination unit determines in advance a recognition target segment to be divided into utterance, by referring to the fixed-phrase grammar.

4. A method for dividing a conversational dialog into utterance by use of conversational dialogs inputted in a plurality of channels, by use of a word database in which spellings and pronunciations of words are described, and by use of a grammar database in which grammar including syntactic rules on words is described, the method comprising the steps of:

detecting a pause location in a channel making a main speech;

detecting an acknowledgement location in a channel not making the main speech;

extracting boundary candidates of the main speech by extracting pauses existing within a predetermined range before and after a base point that is the acknowledgement location; and

outputting a word string of the speech segmented by one of the extracted candidates after dividing the segmented speech into optimal utterance in reference to the word database and grammar database.

5. The method according to claim 4, wherein, in the step of outputting a word string, the likelihoods of speech segments divided by the boundary candidates are calculated in reference to the word database and the grammar database, and then the word string of the speech segment having the highest likelihood is outputted after dividing the speech segment into utterance.

6. The method according to claim 4, wherein the grammar database includes fixed-phrase grammar, acknowledgement grammar and recognition grammar.

7. The method according to claim 6, wherein:

the fixed-phrase grammar includes fixed phrases for starting and ending a confirmation;

the word database includes spellings and pronunciations of the fixed phrases for starting and ending a confirmation.

8. The method according to claim 6, the method further comprising determining a recognition target segment to be divided into utterance by referring to the fixed-phrase grammar.

9. A program embodied in computer memory for dividing a conversational dialog into utterance by use of conversational dialogs inputted in a plurality of channels, by use of a word database in which spellings and pronunciations of words are described, and by use of a grammar database in which grammar including syntactic rules on words is described,

the program causing a computer to execute the functions of:

detecting a pause location in a channel making a main speech;

detecting an acknowledgement location in a channel not making the main speech;

extracting boundary candidates in the main speech by extracting pauses existing within a predetermined range before and after a base point that is the acknowledgement location; and

outputting a word string of the speech segmented by one of the extracted boundary candidates after dividing the segmented speech into optimal utterance in reference to the word database and grammar database.

10. The program according to claim 9, wherein, in the function of outputting a word string, the likelihoods of speech segments divided by the boundary candidates are calculated in reference to the word database and the grammar database, and then the word string of the speech segment having the highest likelihood is outputted after dividing the speech segment into utterance.

11. The program according to claim 9, wherein the grammar database includes fixed-phrase grammar, acknowledgement grammar and recognition grammar.

12. The program according to claim 11, wherein:

the fixed-phrase grammar includes fixed phrases for starting and ending a confirmation;

the word database includes spellings and pronunciations of the fixed phrases for starting and ending a confirmation.

13. The program according to claim 11, the method further comprising a function of determining a recognition target segment to be divided into utterance by referring to the fixed-phrase grammar.